Scheduling Network Resources for Machine Learning Clusters#

Kompella (individual draft, rtgwg)

• Why: Everything else at this meeting reacts to congestion; this draft argues the AI network should be scheduled like compute — a provocative reframe from "detect and signal" to "reserve in advance," with real implications for how TE and schedulers integrate.
• What: A framework proposing that network resources (bandwidth, buffers, queues) be estimated, allocated, and deployed alongside compute for training/inference jobs, with contingency paths — citing SLURM and NVIDIA's KAI Scheduler as the compute-side model and the FANTEL/@Scale/OCP discussions as context.
• How: Scheduling happens at compute-placement time (job size/type imply communication needs) and again when a communication phase starts (CCL API invocations reveal exact endpoints and mode — broadcast, scatter, gather, all-reduce). Uses multipath TE (MPTE) with resource reservation for unicast CCL primitives and multicast TE (MCTE, companion draft-kompella-teas-mcte) for multicast primitives. A key argument: reservation state means transit nodes know the sources, so congestion/event signaling can be sent to the right places (ingresses), and reserved alternates give fast recovery — proactive and reactive mechanisms compose.
• Where applicable: Intra-DC backend networks (multipathing mandatory, unicast + multicast required), inter-DC DCI for distributed training, and metro/WAN edge inference. Early position/framework draft; broadly applicable to anyone running scheduled AI workloads, hypercaler and carrier alike.